When Governments Become Nvidia’s Biggest Buyers

The narrative around Nvidia has shifted from hyperscaler dominance to a quieter but potentially more durable source of demand: the nation-state. In its fiscal 2026 year, sovereign AI — the business of building national-level computing infrastructure for governments — generated $30 billion in revenue. That is more than triple the prior year’s total and accounts for nearly 14 percent of Nvidia’s $216 billion in overall sales. The numbers confirm that state-sponsored AI build-outs are no longer a side experiment; they are a structural pivot in the company’s customer mix.

The catalyst for this pivot was partly geopolitical. US export restrictions on the H20 chip for China forced Nvidia to take a $4.5 billion writedown on inventory and forfeit roughly $8 billion in planned quarterly revenue from that market. Licenses for small H200 shipments to selected Chinese customers have been permitted since February 2026, but not a single dollar has materialized — Chinese authorities continue to block imports at the border. Sovereign AI contracts, from France’s deployment of 18,000 Grace Blackwell systems to the UK’s reservation of significant compute capacity through 2030, have become the deliberate antidote to that lost business.

Nowhere is this build-out more visible than in Europe. At the ISC High Performance 2026 conference in Hamburg, Nvidia unveiled 35 new AI supercomputers spanning 23 countries — a rollout designed to deliver capacity to more than three million researchers. The region has now accumulated 800 exaflops of AI compute since last year, and Nvidia technology powers over 90 percent of all ongoing AI-factory expansions in Europe. Among the biggest installations: the MareNostrum5 upgrade at the Barcelona Supercomputing Center delivers roughly 20 exaflops for training and 33 exaflops for inference; Italy’s IT4LIA system packs 8,000 GPUs and 82 exaflops of training power; BavariaAI’s Blue Swan offers up to 22 exaflops with 1,000 GPUs; and the HLRS HammerHAI in Germany dedicates 850 GPUs to research applications.

Siemens Energy is already putting that firepower to work. Using Nvidia-accelerated simulation, the company slashed compute times for hydrogen gas turbine modeling by as much as 77 percent.

Alongside the European push, Nvidia used the ISC event to showcase the Vera Rubin platform for scientific workloads such as climate modeling. A single Vera Rubin system delivers more than 7 exaflops of AI performance and 5 petaflops of native FP64 support, with up to 144 GPUs per rack and memory bandwidth 2.8 times greater than the previous Blackwell generation. Dell, HPE, Supermicro and other OEMs are planning Vera Rubin-based systems, with availability slated for the fourth quarter of 2026. CEO Jensen Huang has pegged the total addressable market for the Vera CPU architecture at $200 billion, and standalone Vera CPU sales are projected to have already generated $20 billion in revenue by 2026.

Should investors sell immediately? Or is it worth buying Nvidia?

On the robotics front, Nvidia also unveiled Halos for Robotics at the Automate trade show in Chicago — what it calls the first comprehensive safety system for physical AI. Agility Robotics is integrating Halos into its humanoid Digit robots, which are now deployed at Amazon, GXO, Schaeffler and Toyota Motor Manufacturing Canada.

Despite this barrage of announcements, Nvidia’s shares closed 2.5 percent lower on Tuesday at €178.02, about 12 percent below the 52-week high of €202.50 set in May. The stock currently trades just above its 50-day moving average of €180.49, with a relative strength index of 51.2 — neutral territory. The consensus analyst price target stands at €260.83, implying roughly 43 percent upside, a premium that reflects the view that the sovereign AI pivot is not yet fully priced in.

The next major test comes on Wednesday, June 24, 2026, when Nvidia holds its virtual annual shareholder meeting at 18:00 Central European Time. The formal agenda is based on the proxy statement filed May 12, but the real focus will be on management’s commentary around inference monetization — the internal metric Nvidia tracks as “inference equals revenue” for cloud providers and AI developers. Growth in inference workloads would validate that the massive infrastructure build is generating real returns; stagnation would suggest the spending is still speculative.

Investors also have the second-quarter earnings report penciled in for August 26, 2026. Analysts expect revenue of roughly $91.7 billion, up from $81.6 billion in the first quarter. The question is whether the European supercomputer rollouts and sovereign contracts will start showing up in those numbers. Nvidia itself has projected that hyperscaler AI investments alone will hit $1 trillion by 2027. Last quarter, overall revenue grew 85 percent year over year.

Between Blackwell production ramps — still constrained by HBM memory and advanced packaging bottlenecks — and the Vera architecture’s scheduled arrival, Nvidia’s product cycle looks intact. But the structural shift toward government buyers, designed to reduce concentration risk, may ultimately define the next chapter more than any single chip launch. The shareholder meeting will offer the first glimpse of whether that thesis holds.